Household Financial Distress and Voter Participation...By 2016, Republican counties had fully...
Transcript of Household Financial Distress and Voter Participation...By 2016, Republican counties had fully...
Household Financial Distress and Voter Participation
W. Ben McCartney*
This Version: February, 2018Click here for the latest version
JOB MARKET PAPER
Abstract
How does household finance affect the political process? In this paper, I focus on one particular channel of
influence and ask if financially distressed homeowners are more or less likely to participate in elections. To
address this question, I merge deeds records with voter rolls to create a novel panel dataset, and then use
a difference-in-differences design that compares initially highly leveraged homeowners to their equity rich
neighbors and exploits variation in house price declines during the recession. I find that, for highly lever-
aged homeowners, a ten percent decline in local house prices decreases voter participation by two percentage
points. Furthermore, the effects of financial distress are particularly severe for homeowners that live more
than one mile from their polling place, consistent with a resource constraints channel. Back of the enve-
lope calculations suggest that mortgage distress can explain approximately 500,000 abstentions in the 2012
general election.
JEL Classification: D10, D72, H31, R20
Keywords: Household Finance, Financial Distress, Mortgages, Voter Participation, Elections
*Duke University, Fuqua School of Business. Email: [email protected]. I am incredibly grateful to my committee: ManjuPuri (chair), Manuel Adelino, Ronnie Chatterji, John Graham, and David Robinson for their feedback and support. The paper alsobenefited from comments and suggestions from Alon Brav, Anna Cieslak, Sergio Correia, Simon Gervais, Jillian Grennan, David Kaczan,Cam Harvey, Song Ma, Adriano Rampini, Basil Williams, Ming Yang, and seminar participants at the Fuqua Finance Brownbag. Allerrors are my own.
1
1 Introduction
The household finance literature has begun to shed light on the causes and consequences of households’ fi-
nancial decision making, and political economists have long sought to understand the complex interactions
between policies, outcomes, and institutions. A growing body of work explores the connections between these
two fields, documenting and explaining how politics and policies like the mortgage interest deduction, the con-
forming loan limit, and the capital gains tax affect households’ finance and real estate decisions. Here, I aim
to close this loop and ask how household finance feeds back into the political process. Specifically, I explore a
turnout mechanism, especially important in democracies, and test if voters experiencing financial distress are
more or less likely to participate in the political process by voting in elections.
The theoretical predictions are ambiguous (Rosenstone, 1982). On the one hand, financial distress might
increase participation. Voters experiencing distress might seek to effect change or to punish incumbents and
use their vote as a tool for doing so, a hypothesis known as the “angry voter hypothesis.” Newly distressed
households might now believe they have more to gain from their preferred candidate being elected. Or, know-
ing their high debt levels preclude moving, distressed households engage in an effort to make the best of a bad
situation. On the other hand, financially distressed households might be less likely to participate. Financial
distress could cause psychological distress or depression. Households negatively affected by an economic shock
might lose faith in the system and therefore refuse to participate. Finally, financially distressed households,
because they face tighter time and wealth constraints and because voting is not costless, might be less likely
to participate.
Using a novel individual-level dataset and a difference-in-differences style design, I find that financial dis-
tress decreases participation. Specifically, for highly leveraged households, a ten percent drop in house prices
causes a two percentage point decrease in participation. I demonstrate the robustness of this finding by taking
advantage of the richness of the dataset and including an individual-voter fixed effect. Furthermore, I find
that households that live far from their polling places are especially affected by financial distress, consistent
with a resource constraint story.
I use the publicly available North Carolina voter file, an individual level dataset that details, for each of
the more than six million registered voters in the state, their party affiliation and all of the elections they
participated in between 2008 and 2016. Also included are some demographic variables including their age,
race, and gender. Finally, I know their full names and addresses. I use these two identifying pieces of informa-
tion to merge in data from the county deeds registries. These datasets include a list of every property in the
county along with whe owns it, when it was purchased, how much it was purchased for, and how the purchase
was financed. Furthermore, I also know of any refinances that have taken place since and the terms of those
2
loans. Finally, to measure changes in local house prices and construct expected equity positions, I merge in
monthly median zip-code home values from Zillow. With this dataset, I know the name and address of every
registered voter in the state, if they voted, whether they rented or owned and, if they owned, the details of
their outstanding mortgage. I define financial distress as occurring when a household was both initially highly
leveraged and then experienced a severe, negative house price shock.
To identify the effects of financial distress I use a difference-in-differences strategy. For the first difference,
I exploit the varied timing and magnitude of house price declines across zip codes within counties on the
sample of already highly leveraged homeowners (households with loan-to-value ratios above 90% two years
prior to the election). Along with this first difference, I include a battery of control variables and fixed effects.
Specifically, I compare homeowners in the same county, affiliated with the same party, voting in the same
election, whose houses had similar initial values, who moved in during the same year, who made the same
participation choices in the 2008 midterm and general elections, and who are of the same race, age, and gender.
However, the concern is that households living in zip codes where house prices were severe are different, in
meaningful and unobservable ways, from households where house price declines were mild.
To solve this problem, I include a second difference, initial equity position. That is, I compare initially
highly leveraged homeowners to those that were initially equity rich homeowners. These households share
many of those unobservable characteristics that caused them to make the same location decision. They also
share exposure to local macroeconomic conditions like changes in demand, industry shifts, and political adver-
tisement spending. With more equity to cushion the fall in house prices, equity rich homeowners serve as a
compelling placebo group. For example, when highly leveraged households are hit with negative house price
shocks they become unable to cash out refinance, a project they might have been expecting to undertake to
supplement their income. They also, because they are now underwater, become more at risk of foreclosure.
Neither of these conditions are as true for households that were initially equity rich.
A final concern is that location and loan-to-value (LTV) ratio at origination might be both jointly deter-
mined and endogenous to participation responses to home value declines. That is, participation decisions in
the 2008 midterm and 2008 general elections, might not fully control for differences across homeowners. To
solve this remaining issue, I include individual person fixed effects. This fixed effect absorbs all time-invariant
characteristics, even those that are unobservable, driving the location and initial LTV decisions. Using this
strategy means that identification comes just from differences in recent house price declines. I confirm the
main findings. Specifically, the average highly leveraged homeowner is 4 percentage points less likely to vote
after her house price has declined by twenty percent. This key finding supports the hypothesis that financial
distress decreases voter participation.
Next, I explore the economic importance of my findings. Financial distress explains approximately 15,000
3
abstentions in the 2010 general election in North Carolina. In other words, had house prices not declined, voter
participation would have increased by about 1 percentage point through the closing of the distress channel.
In North Carolina, the share of voters treated – by having had initially high LTVs followed by large nega-
tive house price shocks – is similar between Democrats, Republicans, and independent voters. The similar
exposure to the treatment across parties combined with my findings that the effects of house price shocks are
similar across voters in both parties means that election results should not be affected. I confirm this result
finding that, while the effects of decreased house prices on county-level participation are significant, the share
of votes received by each party in North Carolina is unchanged.
Similar effect sizes across the country would mean that more than 500,000 abstentions in the 2012 election
and almost 200,000 abstentions in the 2016 election can be explained by household financial distress. What
this might mean for election results depends entirely on which candidates the financially distressed abstainers
would have voted for. In North Carolina, for example, abstentions due to financial distress were evenly split
between Democratic voters and Republican voters. Further data collection at the voter level will be necessary
to find if this is also the case in the rest of the country or if, in some states, one candidate’s would-be voters were
especially affected. That said, in preliminary county-level analysis, I find that Democratic counties (counties
where Obama received more than 50% of the vote in 2008) saw an average house price decline of 15.6% between
the presidential elections of 2008 and 2012 versus 11.6% in Republican counties. By 2016, Republican counties
had fully recovered with 2008 to 2016 house price growth of 2.3% while Democratic counties 2016 house prices
were still 0.7% below their 2008 levels. At the same time, participation in Democratic counties was 3.7% lower
in 2012 than in 2008 and only 1.6% lower in Republican counties. In 2016, Republican county turnout was
.45% higher than in 2008 but 6.6% lower in Democratic counties. Clearly, these correlations are consistent with
many economic and political stories. But viewed through the lens of this paper’s findings, they are certainly
suggestive.
Finally, I consider the potential channels through which mortgage distress might lower participation. Im-
portantly, the careful identification means we do not have to consider the channels through which financial
distress might increase voter participation since, at least for mortgage distress caused by unexpected de-
clines in house prices, financial distress decreases participation. I consider three channels: one, a time and
wealth constraints channel wherein voters have too few resources to provide for their families, satisfy their
job requirements, and vote (Rosenstone, 1982); two, a psychological distress channel that causes voters to
feel overwhelmed and abandon projects with high real or perceived cognitive costs (Mani et al., 2013); and,
three, a cynicism channel where financial distress decreases the sense of duty people feel to participate in the
democratic process (Riker and Ordeshook, 1968).
To shed some light on the potential mechanism, I test if financial distress has different effects on house-
4
holds who live near their polling places. I find that highly leveraged households within 1 mile of their nearest
polling place are especially affected by falls in house prices. The idea that psychological costs or cynicism
would be different for households living near versus far from their polling place is less intuitive than the idea
that financial distress is relatively worse for households who must drive, park, and stand in a line of unknown
length than for households who can walk to their polling place. Furthermore, surveys from the US Census
find that, in the general elections of 2010, 2012, and 2014 “too busy, conflicting schedule” was the most com-
mon answer given by abstainers when asked why they did not vote1. Overall, a time and wealth constraints
channel, where, for example, financially distressed voters must choose between voting and being late for their
hourly-paid job or voting and not having to pay for childcare that day is the most consistent with the evidence.
However, all channels likely play a role and more work will be necessary to uncover their relative importance.
This paper contributes to the literature examining the role of negative shocks to real estate values on
households. Mian et al. (2013) highlight the role of debt and the importance of household equity in consump-
tion and Baker (2017) further shows that negative income shocks are particularly harmful to households with
high debt to asset ratios. Bernstein (2015) finds that the implicit tax on underwater households, households
who owe more on their mortgage than their home is worth, results in significant decreases to household labor
supply. Those households that continue to work do so for lower wages (Cunningham and Reed, 2013) because
they are less likely to be able to relocate to higher paying jobs (Brown and Matsa, 2016) and, more broadly,
to avoid the double punch of being both underwater and unemployed (Foote et al., 2008). The effects on the
broader economy are also severe, as highly leveraged households are less likely to start firms (Schmalz et al.,
2017) and less likely to successfully pursue innovation projects (Bernstein et al., 2017). Melzer (2017), again
because of the implicit taxes of debt overhang, documents that underwater households cut back on home
improvements. And finally, the health and wellbeing of homeowners also deteriorates because of mortgage
distress (Currie and Tekin, 2015; Deaton, 2012). I contribute evidence that household financial distress also
affects voter participation, a critically important activity for a well-functioning democracy.
I also add to the literature from political science that asks how economic adversity affects voter participa-
tion. These papers focus primarily on unemployment and foreclosure and have yet to reach a consensus on
whether distress increases or decreases voter participation (see, e.g, Burden and Wichowsky, 2014; Cebula,
2017; Estrada-Correa and Johnson, 2012; Hall et al., 2017). Using a novel measure of economic adversity, high
leverage coupled with house price declines, I add convincing evidence in support of the withdrawal hypothe-
sis. My work also contributes to the huge body of work trying to understand why people vote at all and what
affects participation at the margin (for an introduction to this literature see Blais, 2000, 2006; Cancela and
Geys, 2016; Smets and Van Ham, 2013).
1See, e.g., https://www.census.gov/data/tables/2012/demo/voting-and-registration/p20-568.html
5
The implications of my results for policy are as follows. First, policy makers citing the benefits of the
homeownership society (see, e.g, Sodini et al., 2016) should keep in mind that homeownership at any expense
might adversely affect the very outcomes they hope to encourage – voter participation and, more broadly,
civic engagement (Ekman and Amnå, 2012). Similarly, policies limiting LTV ratios might lead to higher voter
participation rates in market downturns (see Cerutti et al. (2017) and DeFusco et al. (2017) for recent papers
discussing these types of policies). Finally, this paper presents a reason to help underwater households modify
their mortgages. Agarwal et al. (2017) show that the Home Affordable Modification Program (HAMP) was
associated with lower rates of foreclosure, milder house price declines, and increases in durable spending.
To their findings, I add novel evidence that HAMP, and programs like it, might also serve to strengthen
communities by halting the declines in voter participation and civic engagement that come with distressed
mortgages.
Finally, my results motivate further work that tries to understand the complex relationships between
household finance, housing policy, election outcomes, inequality, and political institutions. Choices made by
governments have important effects on the real economy and the social welfare of their countries (Atkinson and
Stiglitz, 2015). Citizens have heterogeneous preferences and so democracies hold elections where people can
freely vote for their preferred candidates and policies. If, however, certain groups are less able to participate,
then policy making will fail to reflect the preferences of those unable to vote (Cascio and Washington, 2013;
Chattopadhyay and Duflo, 2004). This paper therefore presents suggestive evidence of a feedback loop between
household financial distress and the growing inequality in the United States – decreased voter participation.
2 Data, Sample Construction, and Summary Statistics
Extant work has used aggregate measures to reach conclusions about the effects of various phenomena on
voter participation and about the effects of financial distress on various outcomes. This strategy is appropri-
ate if the goal is to determine global drivers of global turnout, but is, except under very special circumstances,
inappropriate for identifying causal, person-level economic relationships. Since at least 1950, social science
has known that ecological correlations cannot be used as substitutes for individual correlations (Robinson,
2009). For example, that areas with more unemployment have lower voter turnout does not mean that be-
coming unemployed necessarily decreases an individual’s likelihood of voting. Statistically, this is because the
average within-area individual correlations are not identical to the total individual correlation, as correlations
between independent and dependent variables of interest are generally smaller for relatively homogenous
sub-groups than for the population at large. The conclusion, then, is that making correct inferences about in-
dividual causal effects based on observed aggregate correlations is infeasible. For more theory see King (2013)
6
and for discussions in an empirical setting see, e.g., Arceneaux (2003) and Adelino et al. (2016). In short,
counties do not vote, voters do; and the dataset needs to reflect this.
2.1 Data Sources
The first dataset I use is the North Carolina voter file which covers all registered voters in the state. Made
publicly available by the North Carolina State Board of Elections are lists of all people registered to vote
in every major election. The voter rolls include the name and address, party affiliation, polling place, age,
race, sex, and birth state of each voter. Importantly, I also know if they voted in every local, state, and
national election. I focus on the 2008, 2010, and 2012 general elections and their primaries, to ensure common
eligibility and uniform tops of the ballot for all voters in the state, for a total of six national elections. North
Carolina has a semi-open primary system in which voters registered with a party can vote only in that party’s
primary. Voters who are unaffiliated with a party can vote in either party’s primary. This leads to a high
number of unaffiliated voters in North Carolina relative to the rest of the country. To mitigate this issue, I
assume that voters who always participate in only one party’s primaries are affiliated with that party. I merge
these three datasets together to create a novel panel dataset. With this dataset I know, for every resident of
those counties covered by DataQuick, whether they rented or owned, and, if they owned, how underwater they
found themselves as housing prices fell. I also know, for those homeowners registered to vote, whether or not
they voted.
The second dataset I use sources data from county recorders’ offices related to sale and loan transactions.
These data are also publicly available, but cleaned and published formerly by DataQuick and now by CoreL-
ogic. This dataset covers the near universe of mortgage loans made in North Carolina between 2000 and 2012
and includes information about the borrowers, the lenders, the mortgage, and the securitizing property. Some
counties have no information and the sample is incomplete in early years. But from 2004, all purchase loans
are recorded; and, from 2006 on, all purchase and refinance loans are covered in the most populous counties,
covering about 68% of the state’s population. The third dataset, also from DataQuick, derives its data from
county assessor’s offices and publishes the assessed value, geolocation, street address, and names of owners
for all properties covered in the sample2. To measure local housing market conditions, I use the historical
monthly zip code median home price from Zillow.
North Carolina is a state particularly well suited for this study. The state has ten million citizens and, with
a GDP of approximately $500 billion, an economy just smaller than Norway’s and larger than Venezuela’s.
Also, as shown in Table 1, North Carolina is a remarkably representative state. North Carolina and the2For examples of the raw data, visit the Durham county records search, http://property.spatialest.com/nc/durham/, the
Wake county real estate property search, http://services.wakegov.com/realestate/, and the Wake county register of deeds,http://services.wakegov.com/booksweb/genextsearch.aspx
7
United States are both 61% white. North Carolina has a higher homeownership rate (65.7% versus 63.4%),
lower unemployment rate (4.7% versus 5.0%), and lower median income (54k versus 59k) than the country.
Politically, North Carolina had higher participation in the 2016 general election (64.8% versus 59.3%) and
voted more for Trump (49.8% versus 46.1%). These differences are all slight, though, making North Carolina
an ideal state to use for testing and calibrating the model.
Even in its own right, separately from how it compares to the rest of the country, North Carolina is in-
credibly important. It has been a battleground state since at least the 2008 presidential election when Obama
received just 14,177 more votes than McCain out of 4,310,789 votes cast. Midterm elections, too, are competi-
tive. In the 2014 senatorial race, the Republican Thom Tillis received 45,608 more votes than the incumbent
Democrat Kay Hagan. Most recently, the 2016 governor’s race was decided by just 10,277 votes when Roy
Cooper (D) defeated the incumbent Pat McCrory (R) despite the state voting for Trump.
2.2 Defining Financial Distress
To define financial distress, I start by estimating each household’s monthly loan-to-value (LTV) ratio. To esti-
mate the outstanding loan balance, I assume homeowners pay off their principal in equal monthly payments
over 30 years. I set the value of the home equal to the sale price in the quarter that the sale took place and
then assume it appreciates the same as the median home value in its zip code. The LTV ratio is the quotient
of the two. Then I consider negative shocks to house prices. Finally, I say that highly leveraged homeowners
hit with negative shocks are financially distressed. So, to determine the effects of financial distress I ask what
happens as already highly leveraged households see their home values decline.
I assume that households with low leverage are not likely, or are at least less likely, to be financially
distressed as a result of house price declines. Foote et al. (2008) and Foote et al. (2010) find that falling
house prices and some second negative shock are the key drivers of foreclosure. A decline in house prices
that pushes households underwater is a necessary condition for default. Therefore, households whose home
prices decline become more financially distressed since their risk of default increases. Furthermore, negative
equity might also be financially distressing to households that use their homes as sources of income. When
these households, who were expecting to cash out refinance in the future, learn that they cannot because their
home’s value has declined, they are more at risk of experiencing financial distress.
Motivated by the mortgage distress many households experienced during the recession, the Panel Study of
Income Dynamics (PSID) survey added questions A27F6 and A27G that asked households how likely they are
to fall behind on their mortgage. I use survey responses to questions A20 and A24 to determine if households
are underwater. Of PSID respondents, 37% of underwater homeowners in the 2009 survey say they are worried
8
about falling behind compared to only 11% of homeowners who are not underwater. In 2011, 32% and 8%,
respectively, are concerned. Furthermore, within households over time, being underwater is correlated with a
.6 point increase in K-6 non-specific psychological distress scale, a score out of 24, and a 21 percent increase
in receiving financial help from family and friends (G44). These results are not conclusive evidence, but do
support calling highly leveraged households experiencing negative house price shocks financially distressed.
2.3 Sample Creation
The primary sample creation process is as follows. I start with the North Carolina voter sample which includes,
over the period 2008 to 2012, 8.2 million unique voter-by-address individuals. Given the non-uniqueness of
names (for example, there are six William McCartneys currently registered to vote in North Carolina) I cannot
follow voters as they move across the state, so I use voter-by-address as my panel variable. From the initial
raw dataset, I drop voters living in census tracts not covered by DataQuick, which brings the number of voters
to 5.6 million. I then match voters based on their names and addresses to property owners in the assessor files
and find a match for 59.1% of voters.
I further trim this sample for the regressions by including only those homeowners who were registered to
vote in the 2008 elections. I use their participation in the 2008 midterm and presidential elections as a baseline
off of which to compare their later participation. Because the data quality is questionable pre-2004, I focus on
those properties with observed sales prices between 2004 and 2012. This includes 1.5 million transactions at
just over one million unique properties, or 24% of the properties in the sample. Further, to remove concerns
about incorrectly measured LTV ratios I drop households whose predicted 2011 home value is more than
50% different than its assessed value or who have an outstanding HELOC. Finally, for each election, I drop
homeowners who have refinanced in the previous two years as their LTV is then a function of their refinancing
decision and not the house price shock.
2.4 Summary Statistics
[TABLE 2 HERE]
In table 2, I describe in detail the demographics of the sample of voters registered to vote in the general
elections of 2008, 2010, and 2012. The sample is approximately 45% Democrat and 35% Republican. The
remaining 20% are largely unaffiliated voters, who, because of the semi-open primary in North Carolina,
can vote in either party’s primary. Very close election results suggest that the state is evenly split between
households voting Democrat and households voting Republican. The age, race, and sex variables are largely
unsurprising. Interestingly, only 37% of the registered voters in North Carolina were born in the state. This
9
is due to the large influx of out-of-state immigrants North Carolina has received over the last several decades.
Approximately 60% of the sample are homeowners; the remaining 40% rent. The census estimates that 65% of
North Carolina are owner-occupying homeowners suggesting that my match between the real estate data and
the voter records is working well. Approximately 25% of homeowners in the state live in homes with values
above $300,000. And just under 30% of registered voters live more than one mile away from their nearest
polling place.
[TABLE 3 HERE]
Table 3 looks at the participation rates of registered voters. Democrats and Republicans are far more
likely to vote than unaffiliated voters. Older voters vote more than young voters; homeowners vote more than
renters; and voters who own more expensive homes are more likely to vote. Interestingly, voters that live near
their polling place are not, unconditionally, any less likely to participate in elections with the exception of the
2008 primary.
[FIGURE 1 HERE]
Between 2008 and 2012, house prices fell substantially across the state of North Carolina. Figure 1 illus-
trates the house price falls between the 2008 and 2010 general election in zip codes in Mecklenburg County,
home of Charlotte, North Carolina. In 2008, house prices had been rising almost everywhere for the previous
two years. But by 2010 the trend in the state had reversed and zip codes experienced large, but varied, house
price declines.
[FIGURE 2 HERE]
Figure 2 presents the same zip codes but at four different points in time, May and November of 2010 and
2012, which correspond to the four national elections that took place during my sample period. Immediately
obvious is the two year span leading up to the 2010 elections saw more severe house prices than the two years
leading up to the 2012 elections. And not only did the declines vary in magnitude across zip codes, but also
within zip codes over time. It is this varied magnitude and timing of the falls in house prices that I exploit in
my identification strategy.
[TABLE 4 HERE]
Tables 4 and 5 focus on the mortgage distress variables that will be used in this paper as the key explana-
tory variables of interest. Table 4 documents that the number of highly leveraged households increases from
23.32% of homeowners to 30.94% of homeowners between the 2008 and 2010 elections. This is caused by the
10
large house price declines that rocked the state during that time. Between 2008 and 2010, 51% of voters expe-
rienced local house price declines of more than 5%, and 23% experienced declines of more than 10%. The third
panel groups households into one of six types depending on their initial LTV ratio and the house price drop
they just experienced. For example, at the 2010 election, 5.74% of households had 2008-LTV ratios above 90%
and had just experienced a local house price drop of more than 10% in the two years leading up to the election.
[TABLE 5 HERE]
As a first look at the findings of this paper, simple tabulations find that, in all elections, more leveraged
borrowers are less likely they are to vote. These differences are meaningful, with households with LTVs over
90% being 3 to 12 percentage points less likely to vote than households with LTVs below 90%. Voters living in
zip codes where house price declines had been severe are less likely to participate. And finally, households with
high initial LTV ratios and large declines in home value are less likely to vote – both compared to households
with high initial LTV ratios who did not experience such large house price declines and compared to households
who experienced a similar house price shock.
3 Identification Strategy
The objective of this paper is to uncover the causal relationship between financial distress and voter par-
ticipation. A simple documentation that the participation of households underwater on their mortgages is
lower than households not underwater is intriguing, but nothing more than suggestive. Registered voters who
choose high LTV ratios might be endogenously choosing high LTV ratios and abstention. For example, younger
voters, households with less education, and households with low incomes might all choose more aggressively
financed homeownership and value their vote less or face higher costs to voting than their older, more edu-
cated, or wealthier peers. As another example, homeowners who travel more for work may be both less able
to find time to vote and also have higher LTV ratios because their job means that they relocate often so have
had less time to build equity in their homes. Identifying an effect of financial distress on voter participation
therefore requires some exogenous shock to distress.
To achieve this I use the varied timing and magnitude of house price declines across the state of North
Carolina as a source of exogenous variation to household’s financial distress. I first compare initially highly
leveraged households who experienced large house price declines to other highly leveraged households who
experienced more mild declines in house prices. Then, since households might be non-randomly choosing
neighborhoods such that those in zip codes where house prices fell dramatically might be different in unob-
servable ways from households in zip codes where the recession was mild, I add a second difference to the
11
model. Households with low initial leverage ratios are less likely to become financially distressed when house
prices decline, but have endogenously sorted in to the same neighborhoods and experience the same local
macroeconemic conditions as their highly leveraged neighbors, and thus make a convincing control group.
Specifically, I consider difference-in-differences models of the following form:
Participatedi j = γ×HP Falli j ×High Levi +µ×HP Falli j +θ×High Levi +Controlsi +Election j (1)
where i indexes voters and j indexes elections. Participated is a dummy equal to 100 if the voter participated
and zero otherwise. I use 100 as the outcome so the slope estimates can be easily interpreted in percentage
point terms. The variable of interest, HP Fall, is the percent decrease in zip code house prices over the two
years leading up to the election and High Levi is a dummy variable equal to 1 if household i was highly
leveraged two years prior to the election. The model includes demographic controls about the voter including
his race, sex, age, and birth state; controls for the year he moved in, the value of the home two years prior,
and the percent increase in home prices between the start of 2004 and the end of 2007. Finally, to control
for differences in each individual’s baseline voting participation, I include two indicator variables, the first
equal to 1 if the voter voted in the 2008 midterm election and the second equal to 1 if if he voted in the 2008
presidential election. These participation control variables are important because, if households who chose
to live in areas that experienced house price declines are less likely to participate then this will be controlled
for in the model. The inclusion of election fixed effects, one for each of the 2010 midterm, 2010 general, 2012
midterm, and 2012 general elections, absorbs differences across elections that might be correlated with voter
turnout.
I can further add to the model party affiliation by election fixed effects. This absorbs any differences in
common drivers of voters of different parties to participate in each election. For example, it might be that
voters affiliated with the party out of power are more likely to participate in the midterm elections. If they
also live in zip codes where house price declines were different than zip codes where voters of the other party
live, then the results would be biased. I also include county by election fixed effects to absorb differences
in participation rates across the state and initial house price level by election fixed effects to control for the
concern that voters in wealthier or poorer neighborhoods might have different baseline participation rates or
might be differentially affected by the drop in median local house price. All together then3, this first model
tests for differences between voters experiencing different house price declines but living in the same county,
affiliated with the same party, voting in the same election, living in houses of similar initial value, who made
the same participation choices in the 2008 midterm and general election, and of the same race, age, gender,
3The estimation of models with a high dimensional fixed effects is made possible by Correia (2017).
12
and years since they moved to their home.
A concern, though, is that important unobservable variables might affect both location and participation
decisions. If households choose counties, but not zip codes, zip code level house price changes are randomly
assigned. But to the extent that this is not true, my results will be biased. My first solution to this problem is
to force the identification to come from differences between highly leveraged and low leveraged households in
the same zip code. I do this by estimating model 1 and also including a zip code by election fixed effect. Since
my variation is at the zip code level, the inclusion of this fixed effect naturally means I can no longer identify
the main effect of the house price decline, but I can identify the interaction term between house price falls and
high initial leverage. The assumption is that many of the unobservable characteristics that lead households to
choose the zip codes they did will be shared between high-LTV and low-LTV households. Furthermore, these
households all share exposure to local macroeconomic shocks. The only difference, then, is their initial LTV
ratio and, consequently, how financially distressing a given fall in house prices is. I adjust the strategy further
by including a third group of registered voters, renters. Renters have little exposure to the real estate market,
and are therefore, of the three groups, the least likely to experience financial distress following declines in
house prices.
My second solution acknowledges that households who choose low LTV ratios at origination might be
different than households who put little money down at the time of purchase. Much of this difference will
be absorbed by the participation decisions in the two 2008 elections and the battery of control variables, but
unobserved differences may remain. This is especially true if households make their location and initial
leverage decisions jointly. To fix this issue, I take model 1 and then more fully utilize the panel nature of the
dataset by including a voter fixed effect as follows:
Participatedi j =β×HP Falli j ×High Levi +ρ×HP Falli j +η×High Levi +Personi +Election j (2)
where all variable are as before and Personi is a person fixed effect. This fixed effect absorbs all time-invariant
characteristics, even those that are unobservable, driving the location and origination LTV decisions. In other
words, all the unobserved variables that affect both location decisions and participation decisions that are
not absorbed by the participation decisions in the 2008 elections are absorbed by the person fixed effect.
Each voter has six participation decisions to make, one for each of the national elections occurring during
the sample period. In this model, then, I identify just off of different house price changes the same highly
leveraged homeowner experiences over the years 2006 to 2012. As before, I also include party by election and
county by election fixed effects.
13
4 Results
4.1 The Effects of Financial Distress
[TABLE 6 HERE]
The main results of this paper are presented in tables 6 through 8. In the first specification of table
6, I find that drops in house prices have no significant effect on households that were not initially highly
leveraged. This finding is consistent with the idea that a decline in house prices does not cause equity rich
households to be financially distressed. For households that were initially highly leveraged, on the other
hand, a twenty percent decline in house price decreases voter participation by 3.1 percentage points. With
an average participation rate over the four national elections in 2010 and 2012 of 45 percent for voters in the
regression sample, this decrease is equivalent to a decrease in participation of more than 6 percent. The first
specification includes a battery of control variables for each voter including their race, sex, age, and birth state;
controls for the year they moved in, the value of their home two years prior, and the percent increase in their
zip code’s home values between the start of 2004 and the end of 2007. To control for differences in individual’s
baseline voting participation, I include two indicator variables, the first equal to 1 if the voter voted in the 2008
midterm election and the second equal to 1 if he voted in the 2008 presidential election. Specification (1) also
includes an election fixed effect. As each election is different, this fixed effect is important. Midterm elections
have lower participation levels than presidential elections, and primaries similarly have lower turnout. But
all specifications include an election fixed effect so that correlations between house price declines and the
upcoming election do not affect the results.
Specifications (2) through (5) include increasingly strict fixed effects limiting the sources of variation driv-
ing the identification. Voters affiliated with certain parties might be more inclined to vote in certain elections
than voters affiliated with the opposing party. For example, voters in the minority party might be more in-
clined to participate in the midterms to flip the house and senate. To control for this issue, I replace the
election fixed effect with a party-by-election fixed effect in specification (2). Interestingly, the slope estimates
remain almost completely unchanged, suggesting that differential party participation is not occurring in any
meaningful sense. In the third specification, I add a county-by-election fixed effect to absorb the fact that
some counties might have different mean participation rates. This could be the case if some areas got more
political spending on advertisements or had more important local or congressional races and this was corre-
lated with smaller house price declines. The difference between this specification and the previous ones is
that, while the previous specification compares people to everybody else in the state, this model uses just the
variation between households in the same county. The slope estimates adjust down, consistent with the idea
14
that households sort to counties throughout the state, but the story remains the same. Finally, because houses
of different values were differentially affected by the same local house price drops, I include an initial house
price level-by-election fixed effect. The results consistently point to the same story: that households hit harder
by house price declines are less likely to participate and that this result is not driven by households sorting
into parties or counties or expensive homes.
In the final two specifications, I include zip code fixed effects. Since variation comes at the zip code level,
including the zip code precludes the identification of the main effect of a decrease in house prices. But I can
still compare highly leveraged households to their zip code neighbors who were not initially highly leveraged.
The interaction term estimates are nearly identical to the previous two models, which means that the negative
interaction effect estimated in the previous models was not just picking up differences between people in dif-
ferent zip codes. In these last four specifications, I find that a twenty percent decline in house price decreases
participation for highly leveraged households by about 1.5 percentage points. I present the results of this table
graphically in Figure 3.
[FIGURE 3 HERE]
The chart tells the same story as the tables. The participation rates of households with high leverage
ratios are unaffected if house prices stay the same or increase. And households with low initial leverage ratios
are unaffected by house price declines. But an important effect is immediately obvious at the intersection of
high leverage and large negative shock to house prices. Homeowners in this quadrant, those I call financially
distressed, are significantly less likely to participate in elections.
[TABLE 7 HERE]
Table 7 further compares both highly leveraged homeowners and low initial LTV homeowners to their
renter-neighbors. From specification (3), I find that a twenty percent decline in local house prices decreases a
renter’s participation likelihood by a statistically insignificant .7 percentage points. An initially highly lever-
aged homeowner is a further 1 percentage point less likely to participate while a low LTV homeowner is a
further .68 percentage points less likely. The results from this table say that the more likely a group is to be-
come financially distressed as a result of a negative shock to house prices, the more severe the decrease in voter
participation as a result of a realized negative shock. And that for households that are financially distressed,
house price declines have economically and statistically significant negative effects on participation.
[TABLE 8 HERE]
Table 8 uses a different strategy than the previous models and includes a voter fixed effect. The empirical
literature in political science has included a variety of individual-level factors that might affect turnout (Smets
15
and Van Ham, 2013). By including a voter fixed effect, I control for all of these factors that are time invariant.
The fixed effect also controls for those factors, like career choice and education, that may affect both location
and participation decisions. Including voter fixed effects means that the identification of the estimates comes
just from individual voters whose home value growth changes between elections. That is, there is no concern
that the types of households who were hit hard are also the types of households who do not participate and
that this is driving the results.
I find that, as expected, having a high initial leverage ratio is correlated with lower participation. It is
important to note that this relationship is not causal, as I have no exogenous shock or instrument for initial
leverage ratio. Beyond that, shocks to house prices have dramatic effects on voter turnout, but, again, only on
voters that were initially highly leveraged. The estimated effects of a twenty percent decline in house prices
range from 3.6 to 6.5 percentage points. The fixed effects are included in the same progression as in table
6. Taken together, the results of these three tables say that financial distress, defined as highly leveraged
homeowners experiencing declines in house prices, causes a decline in voter participation.
4.2 Homogeneity of Effects Across Parties
[TABLE 9 HERE]
In table 9 I split households into three types: Democrats, Republicans, and those unaffiliated with any
political party. I then estimate the same models as in table 6 but include another set of interactions. I interact
party affiliation with the financial distress variables and test to see if financial distress has a particularly
strong effect on one party. I find that this is not the case. Between the three political party groups, the effect
of financial distress on participation is statistically indistinguishable.
4.3 Current LTV as a Sufficient Statistic
To further demonstrate the robustness of the findings is to use the current LTV homeowner’s LTV ratio. So
far, I have used previous loan-to-value rations and shocks to house prices as the variables of interest in the
models. But one can instead think of those shocks to house prices as instruments for today’s LTV ratio. I can
then run reduced form regressions using this strategy. Throughout the paper, since I drop from the sample
households who have recently refinanced, changes to household’s LTV ratios are come almost exclusively from
changes to house prices so the two strategies should yield similar results.
[TABLE 10 HERE]
All models include individual voter fixed effects so all identification is coming just off of differences in a
16
voter’s LTV ratio over time. I choose 80% and 100% as cutoffs since they are often used in industry. Households
with LTV ratios above 100% are called underwater and households with LTV ratios between 80% and 100% are
said to have low equity. Some of the negative effects of having low equity are likely endogenous – households
non-randomly choose LTVs ratios at 90% or even 95%. However, it is fair to say that homeowners with LTV
ratios above 100% are financially distressed. Only because of some unexpected, negative shock to home value
can a household become underwater, otherwise the loan would not have been originated. Using the same
fixed effects as in previous tables, I find that, within voters, having low equity makes voters between 1.5 and
2.3 percentage points less likely to participate, and being underwater decreases participation likelihood by a
further 1.3 to 2.6 percentage points. The magnitude of these results are similar to those arrived at using house
price shocks as the explicit source of exogenous variation.
5 The Economic Importance of Distressed Abstentions
In this section, I perform some back of the envelope calculations to show how voter participation is affected by
financial distress. Whether election results are altered depends entirely on which candidates the financially
distressed non-participants would have voted for. I demonstrate that, in North Carolina, would-be Democratic
voters and would-be Republican voters are equally exposed to financial distress. For this reason, election
results are not meaningfully altered in North Carolina. However, I use some county-level data from across
the country to demonstrate that financial distress might play an important role in explaining the 2016 US
Presidential election.
5.1 Implications for North Carolina
I use the model estimates from table 10 along with counts of homeowners to perform some back of the envelope
calculations. From specification (6), households treated with LTV ratios above 100% are 1.5 percentage points
less likely to vote than they would have been had they not been underwater. In the 2010 general election,
9.4% of registered voters were underwater on their homes. Had they instead not been underwater, predicted
participation would have increased by an estimated 8,500 voters. Furthermore, 18% of households had low
equity in their homes. The model predicts that had all households with low or negative equity had LTV ratios
below 80%, then 36,313 more voters would have participated. In other words, voter participation in the 2010
midterm election would have been .3 to 1.3 percentage point higher in the absence of financial distress.
These estimates are partial equilibrium results estimated with a latent variable model. To truly explain
abstentions requires some sort a structural or general equilibrium model. The exercise here is meant only to
illustrate the potentially important effects mortgage distress has on voter participation and election results.
17
To know how the election results would have differed had these non-participants voted requires knowing
for whom they would have voted. This is impossible to know with certainty, but using the data to compare
political affiliations of participants with election results allows me to make some good guesses. Some results
are available at the precinct level, but, in almost all cases, early and absentee ballots are tabulated only at
the county level. For this reason, I use county level results. I find that the share of households registered as
Democrats is highly correlated with the share of the county that voted for the Democratic county across all
elections.
[FIGURE 4 HERE]
Knowing that observed party affiliation is highly correlated with realized vote, at least at the aggregated
county level, means that we can assume who a voter would have voted for had she participated. The next step
is to uncover if Democrats or Republicans were more likely to have experienced financial distress due to high
initial LTV ratios and then negative shocks to house prices.
[TABLE 11 HERE]
In table 11, I document that the share of households experiencing the double whammy of high initial
leverage and large house price declines is similar between Democrats, Republicans, and Independents. In
fact, the share of each party in each of the six treatment groups is remarkable similar. To provide further
evidence showing that exposure to financial distress was similar across parties I present figure 5.
[FIGURE 5 HERE]
Figure 5 demonstrates that, even within counties, the share of Democrats being shocked into mortgage
distress is highly correlated with the share of Republicans experiencing the same fate. In other words, it is not
the case that in some counties all the Democrats were hit hard by a house price shock and the Republicans
were all spared. If a certain share of the county’s Democrats were affected, a similar share of the county’s
Republicans was as well. I also check to see if there is a relationship between how Democratic a county is
and the house price declines it experienced leading up to the 2016 election. I find that the relationship is
flat, meaning that house price falls did not occur just in Democratic counties or just in Republican counties.
Taken together, these findings suggest what we already knew from table 11, that voters of both parties were
similarly exposed to negative house price shocks and similarly at risk of financial distress as a result. Recall,
importantly, that the effects of mortgage distress are similar on Democrats, Republicans, and Independents.
Given the similar exposure to the shock and the similar effect of the shock across parties, the results in table
12 are exactly as expected.
18
[TABLE 12 HERE]
The first three models, regressing change in turnout on house price declines, confirm that voter participa-
tion goes down following declines in house prices. But, in models 4 and 5, I document that the share of the
county’s vote for a given party does not change. In fact, including the county FE, notably without an election
fixed effect, explains almost everything there is to explain in the county’s collective vote. It is important to
understand that this is because, within every county, the share of Democrats affected was the same as the
share of Republicans affected. In other words, if there was no financial distress, the number of voters would
increase, but the share of these counterfactual distress-free voters voting Democrat would be the same as the
share of voters voting Democrat in the presence of the house price drops. This is nothing mechanical, but
purely a function of the locations of party members across the state of North Carolina and where house price
drops occurred.
5.2 Implications for the United States
To perform the same analysis across the United States requires household level data of party affiliations. For
now, though, I can use a similar methodology as before to estimate abstentions. CoreLogic’s Equity Report
form the fourth quarter of 20164 publishes that 15% of mortgaged residential properties have LTV ratios
between 80% and 100% and a further 6.2% are underwater. In the United States at the time of the election,
approximately 130 million homeowners had mortgages. Using these numbers along with the assumption
that the estimates from table 10 are appropriate for the rest of the country, I calculate that between 117,000
and 590,000 abstentions were because of mortgage distress. Performing the same analysis in 2012 when
21.5% of homeowners were underwater and another 23% had low equity5 I estimate financial distress caused
between 407,000 and 1,400,000 abstentions. Again, I urge caution as estimates are partial equilibrium results
estimated with a latent variable model. These calculations are merely meant to be illustrative.
Determining how election results in other states might have been affected in the absence of house price
shocks and mortgage distress requires more data. Specifically, I would need to know, or least be confident
guessing, for whom the financially distressed abstainers would have voted for had they participated. In North
Carolina I know each voter’s political affiliation, so I can do this. I do not yet have this information for
other states. Instead, I create a simple dataset I use to present some suggestive results. For each county in
the country where Zillow data is available, I know the percent change in median house price that the county
experienced between the 2008 and 2016 elections. I also know the share of the county vote that went to Obama
4http://www.corelogic.com/research/negative-equity/equity-report-q4-2016-screen-030817.pdf5http://www.corelogic.com/research/negative-equity/corelogic-q4-2012-negative-equity-report.pdf
19
in 2008, which I use as a measure for how Democratic the county is. I then calculate a simple correlation
between how democratic a county was in 2008 and how their home values have changed since then.
I find that Democratic counties (counties where Obama received more than 50% of the vote in 2008) saw
an average house price decline of 15.6% between the presidential elections of 2008 and 2012 versus 11.6% in
Republican counties. By 2016, Republican counties had fully recovered with 2008 to 2016 house price growth
of 2.3% while Democratic counties 2016 house prices were still 0.7% below their 2008 levels. At the same time,
participation in Democratic counties was 3.7% lower in 2012 than in 2008 and only 1.6% lower in Republican
counties. In 2016, Republican county turnout was .45% higher than in 2008 but 6.6% lower in Democratic
counties. Clearly, these correlations are consistent with many economic and political stories. But viewed
through the lens of this paper’s findings, they are certainly suggestive. These results can also be presented
graphically.
[FIGURE 6 HERE]
In figure 6, I present these correlations for four different states. For North Carolina, I find a gently upward
sloping trend meaning that areas that were more Democratic in 2008 had larger house price growth between
2008 and 2016. The size of the point corresponds to the population of the county. The other three states I
look at, Michigan, Pennsylvania, and Wisconsin, were all carried, unexpectedly, by Trump. In each state,
there are large, Democratic counties where house prices are still significantly below their 2008 levels. The
suggestive result then is that had these counties recovered, fewer, likely Democratic households, would have
been financially distressed. Trump won Michigan, Pennsylvania, and Wisconsin by just 10,704, 44,292, and
22,748 votes, respectively. So the effects of financial distress need not have been unrealistically severe to have
had a dramatic effect on the outcome.
6 Channels
6.1 The Potential Channels
In this paper, using multiple identification strategies and a carefully constructed, detailed dataset, I provide
evidence in support of the hypothesis that financially distressed homeowners are less likely to participate in
elections. The clean identification and unambiguous results provide an important piece of evidence in the
puzzle of understanding why people vote and what affects their participation. I now consider some of the
channels, discussed in previous sections, through which financial distress might decrease participation.
The first channel, and the channel most consistent with my evidence, is a financial or time constraints
mechanism. This mechanism operates on households who want to participate but are too constrained to do so.
20
Consider a few examples. A registered voter wants to participate, but because of the decline in house prices
she was not able to refinance as she had planned to do. To avoid foreclosure, she takes a second job. Between
her two jobs, she does not have time to make it to the polls and therefore does not participate. In another
scenario, a household had been paying for child care. But, following the house price declines and the larger
threat of foreclosure, the household has adjusted their schedule to take care of their own children and can no
longer find the time to participate. A third story is of a low-income voter who works an hourly job with has
a boss who will not allow him to take time off or come in late. His financial constraint has no slack and he
can’t afford to forgo the hour of work and the increased threat in being laid off. In all cases, it is important
to remember that voting is not a quick activity that occurs just on election day. Voters must register to vote,
which might not be trivial especially for households without internet access, make sure they know where to
vote, and then get to the polling place and potentially stand in line for hours. To a household that is operating
with little slack to their budget and time constraints, this project might be too costly to undertake.
A second potential channel is the psychological distress channel. This mechanism causes the results if it is
the case that households experiencing financial constraints experience a cognitive overload. If their capacity
to work through complex problems is negatively affected because of their financial distress they may drop
potential projects from their to-do lists. In this case, it is not that they lack the resources to undertake the
project, but that the psychological stress of being financially distressed impairs their cognitive function. A
third channel is that financial distress affects a household’s perceived benefit of voting. Households experienc-
ing financial distress might observe their local economy collapsing and wonder if their vote can even matter.
I call this the disillusionment or cynicism channel. Many other stories, motivated by the models discussed in
the appendix, can also be told.
6.2 Evidence in favor of the Constraints Channel
To speak to potential channels, I present three pieces of evidence. The first is that households are less likely
to vote, not just because of house price declines, but because of house price declines when initially highly
leveraged. If the disillusionment channel was the key channel, then we would expect to see that renters
and households living in areas where house prices fell dramatically would be equally affected. It is unclear
that a collapsing economy would increase cynicism just for households that were initially highly leveraged. I
therefore rule out this channel as having first order importance.
Next, I use the voter’s distance to his nearest polling place to measure his cost of voting since, all else
equal, being farther away from the polling place makes it costlier to vote. If it was a case of jadedness with
the system or mental anxiety, I would not expect the interaction between financial distress and distance to
21
matter. If, however, it is especially costly to vote when time and resources become relatively more valuable,
then a negative interaction effect financial distress and faraway polling place is most consistent with a time
and wealth constraints channel.
[TABLE 13 HERE]
In table 13 I compare households whose nearest polling place is more than one mile away with households
whose polling place is less than one mile away. I find that this triple interaction is economically and statisti-
cally significant. In other words, highly leveraged households who are also more than a mile away from their
polling place are especially affected by falls in house prices. To ensure that this is not just a feature of differ-
ences between areas where there are many polling places and area with few polling places I include a polling
place by election fixed effect. The results are unchanged. I interpret these results as saying that to voters for
whom voting, as a function of travel time and unknown line length, is relatively costless, financial distress
imposes less of a burden. If it was the case that households chose to abstain because of disillusionment or
psychological distress and distraction then it should not be the case that being near to the polling place would
matter in this way. That said, one might argue that households who live near to the polling place are psycho-
logically distressed, it is just that since they can walk to their polling place the psychological distress is less
relevant than for households for whom voting is a larger imposition. To provide a test that can more cleanly
disentangle the financial constraints channel from the psychological distress channel I consider one final piece
of evidence.
After every election, the US census polls non-participants and asks why they did not participate6. In the
2016 general election, the third most given reason for not participating, at 14.3 percent, was that households
were too busy or had a conflicting schedule. And time constraints was the most cited reason after the 2014,
2012, and 2010 general elections at 28% 19%, and 27%, respectively. Differences across the income groups
also show that for households with low incomes, transportation problems and forgetting to vote are more se-
vere problems than for wealthier survey respondents. That so many households cite time constraints as an
insurmountable hurdle suggests that time and financial constraints play a relatively larger role in financial
distress decreasing participation than psychological distress. That all said, perhaps financial distress and psy-
chological distress occur together and psychological distress causes households to feel overwhelmed and "too
busy". A more refined approach will be necessary to completely disentangle the psychological and constraints
channels, but I view the evidence as most consistent with the latter.
6https://www.census.gov/data/tables/time-series/demo/voting-and-registration/p20-580.html
22
7 Conclusion
The key contribution of this paper is a clean identification of the effects of household financial distress on
voter participation. Voting is an activity at the epicenter of all democracies and it is consequently important
to understand how households choose whether or not to participate. I find that distress significantly decreases
participation. To my knowledge, this is the first paper to document this relationship. The aggregate effects are
large, explaining hundreds of thousands of abstentions each election. This paper deepens our understanding
of how household finance affects the political process and calls for more work to be done.
23
ReferencesAdelino, M., Schoar, A. and Severino, F. (2016), ‘Loan originations and defaults in the mortgage crisis: The
role of the middle class’, The Review of Financial Studies 29(7), 1635–1670.
Agarwal, S., Amromin, G., Ben-David, I., Chomsisengphet, S., Piskorski, T. and Seru, A. (2017), ‘Policy inter-vention in debt renegotiation: Evidence from the home affordable modification program’, Journal of PoliticalEconomy 125(3), 654–712.
Aldrich, J. H. (1993), ‘Rational choice and turnout’, American journal of political science pp. 246–278.
Arceneaux, K. (2003), ‘The conditional impact of blame attribution on the relationship between economicadversity and turnout’, Political Research Quarterly 56(1), 67–75.
Atkinson, A. B. and Stiglitz, J. E. (2015), Lectures on public economics, Princeton University Press.
Baker, S. (2017), ‘Debt and the response to household income shocks: Validation and application of linkedfinancial account data’, Journal of Political Economy .
Bernstein, A. (2015), ‘Household debt overhang and labor supply’.
Bernstein, S., McQuade, T. and Townsend, R. R. (2017), ‘Does economic insecurity affect employee innovation?’.
Blais, A. (2000), To vote or not to vote?: The merits and limits of rational choice theory, University of PittsburghPre.
Blais, A. (2006), ‘What affects voter turnout?’, Annu. Rev. Polit. Sci. 9, 111–125.
Brown, J. and Matsa, D. A. (2016), Locked in by leverage: Job search during the housing crisis, Technicalreport, National Bureau of Economic Research.
Burden, B. C. and Wichowsky, A. (2014), ‘Economic discontent as a mobilizer: unemployment and voterturnout’, The Journal of Politics 76(4), 887–898.
Cancela, J. and Geys, B. (2016), ‘Explaining voter turnout: A meta-analysis of national and subnational elec-tions’, Electoral Studies 42, 264–275.
Cascio, E. U. and Washington, E. (2013), ‘Valuing the vote: The redistribution of voting rights and state fundsfollowing the voting rights act of 1965’, The Quarterly Journal of Economics 129(1), 379–433.
Cebula, R. J. (2017), ‘Unemployment and voter turnout revisited: A brief note’, Electoral Studies 48, 149–152.
Cerutti, E., Claessens, S. and Laeven, L. (2017), ‘The use and effectiveness of macroprudential policies: newevidence’, Journal of Financial Stability 28, 203–224.
Chattopadhyay, R. and Duflo, E. (2004), ‘Women as policy makers: Evidence from a randomized policy experi-ment in india’, Econometrica 72(5), 1409–1443.
Correia, S. (2017), Linear models with high-dimensional fixed effects: An efficient and feasible estimator,Technical report. Working Paper.
Cunningham, C. and Reed, R. R. (2013), ‘Negative equity and wages’, Regional Science and Urban Economics43(6), 841–849.
Currie, J. and Tekin, E. (2015), ‘Is there a link between foreclosure and health?’, American Economic Journal:Economic Policy 7(1), 63–94.
Deaton, A. (2012), ‘The financial crisis and the well-being of americans 2011 oep hicks lecture’, Oxford economicpapers 64(1), 1–26.
DeFusco, A. A., Johnson, S. and Mondragon, J. (2017), ‘Regulating household leverage’.
24
Downs, A. (1957), ‘An economic theory of political action in a democracy’, Journal of Political Economy65(2), 135–150.
Ekman, J. and Amnå, E. (2012), ‘Political participation and civic engagement: Towards a new typology’, Hu-man affairs 22(3), 283–300.
Estrada-Correa, V. and Johnson, M. (2012), ‘Foreclosure depresses voter turnout: neighborhood disruption andthe 2008 presidential election in california’, Social Science Quarterly 93(3), 559–576.
Ferejohn, J. A. and Fiorina, M. P. (1974), ‘The paradox of not voting: A decision theoretic analysis’, Americanpolitical science review 68(2), 525–536.
Foote, C., Gerardi, K., Goette, L. and Willen, P. (2010), ‘Reducing foreclosures: No easy answers’, NBERMacroeconomics Annual 24(1), 89–138.
Foote, C. L., Gerardi, K. and Willen, P. S. (2008), ‘Negative equity and foreclosure: Theory and evidence’,Journal of Urban Economics 64(2), 234–245.
Fowler, J. H. (2006), ‘Altruism and turnout’, The Journal of Politics 68(3), 674–683.
Gerber, A. S., Green, D. P. and Shachar, R. (2003), ‘Voting may be habit-forming: evidence from a randomizedfield experiment’, American Journal of Political Science 47(3), 540–550.
Geys, B. (2006), ‘Explaining voter turnout: A review of aggregate-level research’, Electoral studies 25(4), 637–663.
Hall, A. B., Yoder, J. and Karandikar, N. (2017), ‘Economic distress and voting: Evidence from the subprimemortgage crisis’.
Jackman, R. W. (1987), ‘Political institutions and voter turnout in the industrial democracies’, American Polit-ical Science Review 81(2), 405–423.
King, G. (2013), A solution to the ecological inference problem: Reconstructing individual behavior from aggre-gate data, Princeton University Press.
Lipsman, N. and Lozano, A. M. (2011), ‘The most cited works in major depression: the ‘citation classics”,Journal of affective disorders 134(1), 39–44.
Mani, A., Mullainathan, S., Shafir, E. and Zhao, J. (2013), ‘Poverty impedes cognitive function’, Science341(6149), 976–980.
McCartney, W. B. and Shah, A. M. (2016), ‘I’ll have what she’s having: Identifying social influence in householdmortgage decisions’.
Melzer, B. T. (2017), ‘Mortgage debt overhang: Reduced investment by homeowners at risk of default’, TheJournal of Finance 72(2), 575–612.
Mian, A., Rao, K. and Sufi, A. (2013), ‘Household balance sheets, consumption, and the economic slump’, TheQuarterly Journal of Economics 128(4), 1687–1726.
Mueller, D. C. D. C. (1989), Public choice ii; a revised edition of public choice, Technical report.
Niemi, R. G. (1976), ‘Costs of voting and nonvoting’, Public Choice 27(1), 115–119.
Plutzer, E. (2002), ‘Becoming a habitual voter: Inertia, resources, and growth in young adulthood’, Americanpolitical science review 96(1), 41–56.
Riker, W. H. and Ordeshook, P. C. (1968), ‘A theory of the calculus of voting’, American Political Science Review62(1), 25–42.
Robinson, W. S. (2009), ‘Ecological correlations and the behavior of individuals’, International journal of epi-demiology 38(2), 337–341.
25
Rosenstone, S. J. (1982), ‘Economic adversity and voter turnout’, American Journal of Political Science pp. 25–46.
Rosenstone, S. J. H., Rosenstone, J. M. J. and Hansen, J. M. (1993), Mobilization, participation, and democracyin America, Macmillan Publishing Company,.
Schmalz, M. C., Sraer, D. A. and Thesmar, D. (2017), ‘Housing collateral and entrepreneurship’, The Journalof Finance 72(1), 99–132.
Smets, K. and Van Ham, C. (2013), ‘The embarrassment of riches? a meta-analysis of individual-level researchon voter turnout’, Electoral Studies 32(2), 344–359.
Sodini, P., Van Nieuwerburgh, S., Vestman, R. and von Lilienfeld-Toal, U. (2016), Identifying the benefits fromhome ownership: A swedish experiment, Technical report, National Bureau of Economic Research.
Verba, S. and Nie, N. H. (1972), Participation in America, Harper & Row.
26
Figure 1: Percent Drop in Charlotte, NC Zip Code House Prices
This is a map of zip codes in Mecklenburg County, NC as of November of 2010. For each zip code, I compute the change in the median house price overthe previous 24 months using data from Zillow. The larger the decline in house prices, the more heavily shaded the zip code.
27
Figure 2: Percent Drop in Charlotte, NC Zip Code House Prices
This is four maps of Mecklenburg County, NC at four different points of time. The second panel is identical to figure 1. For each zip code at each election,I compute the change in the median house price over the previous 24 months using data from Zillow. The larger the decline in house prices, the moreheavily shaded the zip code.
28
Figure 3: Percentage Point Change in Participation Likelihood as a Function of Initial LTV Ratio and Drop in House Prices
To create this figure, I first regressed the voter’s participation decision on his initial LTV, the decline in house prices he experienced over the previoustwo years, and the interaction of these two continuous variables. Included in this figure are the controls and fixed effects corresponding to specification(3) in table 6. I then graph the marginal effects of different changes in initial LTV ratio and house price declines.
29
Figure 4: Votes Received by the Democratic Nominee over Democratic Voters
This chart presents county level data. For each county and each presidential election between 2008 and 2016, I calculate the share of the county’sDemocrat-affiliated voters that participated and the share of the counties votes that went to the Democratic nominee for president. If voters affiliatedwith the Democratic party vote for the Democratic nominee and Republican and independent voters vote for other nominees, then these two sharesshould be the same. To make the comparison easier, I also graph a 45-degree line and linear trendlines for each election.
30
Figure 5: Percent Increase in Share of Households Underwater Between 2008 and 2012
This is a chart where each point corresponds to each county. For each county I calculate the percent increase in the share of Democrats underwateron their mortgages and the percent increase in the share of Republicans underwater on their mortgages. If voters affiliated with the two parties weresimilarly affected by declines in house prices, then we would expect these two percent increases to be the same. To make the comparison easier, I alsograph a 45-degree line and a linear trendline.
31
Figure 6: How House Prices Changed in Democratic Counties Across Four States
Below are four charts, one each for four different states – Michigan (MI), North Carolina (NC), Pennsylvania (PA), and Wisconsin (WI). Each circlecorresponds to a county in the state and the larger the circle the bigger the population of the county. The placement of the circle on the chart isdetermined by two things, the share of the county that voter for Obama in 2008 and the percent change in house prices between the 2008 and 2016general elections.
32
Table 1: Comparing North Carolina to the Other States
I use data from several sources, including electproject.org, the Bureau of Labor Statistics, and the US CensusCurrent Population Survey. I rank states from highest share voting for Trump to lowest share voting forTrump. In bold are the rows for North Carolina, the sample state used in this paper, and the United States.
State Eligible 2016 Share Won Median Unemployment White HomeownershipVoters Participation by Trump Income Rate
WV 1,425,962 50.10% 68.5% 44,354 5.8% 93.0% 74.8%WY 428,283 59.70% 67.4% 57,829 5.4% 86.0% 70.2%OK 2,773,970 52.40% 65.3% 50,943 5.3% 64.0% 66.8%ND 565,031 60.90% 63.0% 60,184 3.1% 85.0% 61.4%KY 3,276,651 58.70% 62.5% 45,369 5.0% 84.0% 67.9%AL 3,601,361 59.00% 62.1% 47,221 5.4% 66.0% 69.7%SD 632,989 58.50% 61.5% 57,450 2.9% 82.0% 69.4%TN 4,899,384 51.20% 60.7% 51,344 4.6% 73.0% 66.4%AR 2,142,571 52.80% 60.6% 45,907 4.0% 72.0% 67.6%ID 1,167,200 59.10% 59.3% 56,564 3.8% 82.0% 70.5%NE 1,349,903 62.50% 58.8% 59,374 3.2% 78.0% 68.0%LA 3,380,951 60.00% 58.1% 42,196 6.4% 59.0% 64.2%MS 2,176,312 55.60% 57.9% 41,099 6.0% 58.0% 69.7%IN 4,852,657 56.40% 56.8% 56,094 4.5% 81.0% 70.9%MO 4,511,812 62.30% 56.8% 55,016 5.2% 79.0% 66.7%KS 2,051,750 57.70% 56.7% 56,810 4.4% 74.0% 67.1%MT 804,381 61.80% 56.2% 57,075 4.3% 89.0% 67.1%SC 3,706,769 56.70% 54.9% 54,336 4.9% 65.0% 68.9%TX 17,396,296 51.60% 52.2% 58,146 4.8% 43.0% 61.5%OH 8,737,173 62.90% 51.7% 53,985 4.8% 78.0% 66.1%AK 519,849 61.30% 51.3% 75,723 6.8% 60.0% 65.2%IA 2,290,215 68.40% 51.2% 59,094 4.2% 86.0% 70.0%GA 6,955,436 59.20% 50.8% 53,527 5.1% 52.0% 62.3%NC 7,318,442 64.80% 49.8% 53,764 4.7% 61.0% 65.7%FL 14,572,210 64.60% 49.0% 51,176 4.7% 55.0% 64.3%AZ 4,734,313 55.00% 48.7% 57,100 5.5% 54.0% 61.9%PA 9,701,644 63.60% 48.2% 60,979 5.7% 76.0% 68.5%MI 7,423,233 64.70% 47.5% 57,091 4.6% 75.0% 72.8%WI 4,288,320 69.40% 47.2% 59,817 4.1% 80.0% 67.7%NH 1,042,102 71.40% 46.6% 76,260 2.9% 92.0% 71.8%USA 230,585,915 59.30% 46.1% 59,039 5.0% 61.0% 63.4%UT 1,995,987 56.70% 45.5% 67,481 3.4% 79.0% 71.3%NV 1,964,097 57.30% 45.5% 55,431 5.8% 52.0% 54.5%MN 3,966,155 74.20% 44.9% 70,218 4.0% 80.0% 72.4%ME 1,060,905 70.50% 44.9% 50,856 4.1% 92.0% 72.6%VA 6,027,262 66.10% 44.4% 66,451 4.0% 61.0% 66.3%CO 3,966,297 70.10% 43.3% 70,566 3.6% 70.0% 62.4%DE 689,125 64.40% 41.7% 58,046 4.3% 62.0% 73.0%NJ 6,042,792 64.10% 41.0% 68,468 5.3% 58.0% 62.2%CT 2,561,555 64.20% 40.9% 75,923 5.4% 67.0% 64.2%NM 1,456,551 54.80% 40.0% 48,451 6.7% 37.0% 67.4%OR 3,012,502 66.40% 39.1% 59,135 5.5% 75.0% 62.6%RI 786,033 59.00% 38.9% 61,528 5.6% 72.0% 56.3%IL 8,943,045 61.90% 38.8% 61,386 5.5% 61.0% 65.3%WA 5,121,782 64.80% 36.8% 70,310 5.6% 66.0% 61.6%NY 13,591,250 56.80% 36.5% 61,437 5.0% 57.0% 51.5%MD 4,176,484 66.60% 33.9% 73,760 4.2% 53.0% 66.5%MA 4,947,241 67.20% 32.8% 72,266 3.6% 72.0% 59.7%CA 25,017,408 56.70% 31.6% 66,637 5.5% 38.0% 53.8%VT 494,879 63.70% 30.3% 60,837 3.3% 94.0% 71.3%HI 1,016,971 42.20% 30.0% 72,133 3.3% 19.0% 57.7%DC 511,463 60.90% 4.1% 70,982 6.1% 38.0% 40.8%
33
Table 2: Demographics of the Registered NC Voters in the Sample
This table presents summary statistics on the sample of registered voters in each of the three general electionsin the sample. Democrats and Republicans are defined as such if they are either registered with that partyor have only voted in that party’s primary and never the other party’s. All other voters are classified asindependents. Birth year is the year the voter was born. Born in North Carolina is a dummy variable equalto 1 if the voter was born in the state of North Carolina and 0 otherwise. Homeowners are defined as such ifthey are in the DataQuick sample as owners of the home they live in and this address matches their addressin the voter file. I also call other registered voters who live in the same house as an owner-occupant sincethis group is largely spouses not on the deed. Polling place data exist for 60% of the sample and the share ofvoters in each bucket is out of that 60% of the sample. The demographic variables come from the voter file.The homeowner’s home value is from DataQuick.
Election
2008 General 2010 General 2012 General
Party AffiliationIndependent 17.99% 18.39% 19.02%Democrat 46.05% 45.45% 45.42%Republican 35.96% 36.16% 35.56%
Birth Year1942 and Prior 13.94% 12.69% 10.83%1943 - 1958 24.17% 24.05% 22.98%1959 - 1974 29.85% 29.60% 28.45%1975 - 1990 23.84% 23.56% 23.26%1991 and Later 1.12% 2.87% 7.04%
Race, Sex, Birth PlaceWhite 69.82% 69.64% 67.51%African American 18.74% 18.52% 19.45%Hispanic or Latino 1.14% 1.28% 1.67%Male 41.99% 41.99% 41.82%Born in North Carolina 37.50% 37.69% 37.48%
HomeownershipHomeowner 60.13% 59.16% 58.02%
Homeowner’s Home Values$0 - $125,000 19.32% 22.06% 23.36%$125,001 - $200,000 30.71% 30.98% 30.53%$200,001 - $300,000 23.73% 22.92% 22.62%$300,001 + 26.25% 24.04% 23.49%
Nearest Polling Place< 0.5 Miles Away 38.17% 37.29% 37.55%Between 0.5 and 1 Miles Away 33.12% 33.07% 32.96%More than 1 Mile Away 28.71% 29.64% 29.49%
Number Registered 3,877,929 3,856,036 4,048,668
34
Table 3: Participation Rates of Registered Voters
For each demographic group I determine the share of registered voters in the group that participated in each of the national elections between 2008 and2012. All variables are defined as in Table 2.
Election
2008 Primary 2008 General 2010 Primary 2010 General 2012 Primary 2012 General
Party AffiliationIndependent 23.01% 57.53% 8.31% 25.01% 25.78% 59.77%Democrat 54.03% 73.81% 11.14% 38.40% 38.34% 76.04%Republican 26.05% 74.25% 16.10% 45.51% 46.06% 78.92%
Birth Year1942 and Prior 51.10% 74.59% 25.99% 57.40% 52.05% 78.69%1943 - 1958 48.56% 80.03% 19.90% 56.27% 51.59% 83.86%1959 - 1974 35.20% 72.46% 9.82% 39.38% 38.02% 77.10%1975 - 1990 22.24% 56.83% 3.03% 16.56% 25.22% 62.18%1991 and Later 28.41% 62.10% 1.18% 4.86% 19.41% 56.75%
Race, Sex, Birth PlaceWhite 35.19% 70.85% 14.35% 41.43% 43.17% 74.25%African American 53.06% 72.33% 9.33% 36.52% 29.08% 75.31%Hispanic 24.91% 58.50% 2.01% 14.51% 15.84% 58.87%Male 36.31% 70.14% 13.51% 40.22% 38.61% 72.69%Female 39.72% 71.69% 11.64% 37.27% 39.04% 74.89%Born in North Carolina 39.23% 71.46% 13.77% 39.94% 39.39% 74.00%Born Outside North Carolina 37.69% 70.79% 11.60% 37.64% 38.52% 73.95%
HomeownershipHomeowner 44.10% 79.11% 16.84% 50.96% 47.35% 82.02%Renter 28.65% 58.87% 5.73% 20.47% 25.73% 62.84%
Homeowner’s Home Values$0 - $125,000 38.20% 73.03% 11.32% 38.38% 40.46% 78.29%$125,001 - $200,000 38.59% 75.70% 9.28% 38.68% 43.09% 81.34%$200,001 - $300,000 39.53% 79.50% 10.78% 43.52% 47.38% 85.39%$300,001 + 39.91% 82.58% 12.07% 47.35% 47.74% 88.10%
Nearest Polling Place< 0.5 Miles Away 40.48% 68.84% 11.26% 35.11% 38.56% 73.56%Between 0.5 and 1 Miles Away 38.46% 70.50% 11.24% 37.35% 38.33% 74.77%More than 1 Mile Away 34.59% 71.58% 11.98% 38.95% 38.19% 74.29%
35
Table 4: House Price Changes of Registered Voters
This table presents summary statistics on the sample of registered voters in each of the three general elections in the sample. The construction of theloan-to-value (LTV) ratios is described in detail in the Data section of the paper but, in short, is the ratio of the outstanding loan amount to the value ofthe home securing the loan. The percent drop in house prices is the drop in the zip code’s median home price. Note that only homeowner’s have a definedLTV Ratio and Treatment Group, but all registered voters experienced some change in median zip code house price.
Election
2008 General 2010 General 2012 General
Homeowner’s LTV RatioLTV less than 90% 76.68% 69.06% 66.74%LTV greater than 90% 23.32% 30.94% 33.26%
Zip Code House Price FallsFall in House Prices < 5% 97.39% 48.49% 81.44%5% < Fall in House Prices < 10% 2.25% 28.42% 16.67%10% < Fall in House Prices 0.36% 23.09% 1.89%
Initial LTV by Fall in House PricesInitial LTV Ratio < 90% × Fall in House Prices < 5% 81.48% 37.36% 57.94%Initial LTV Ratio < 90% × 5% < Fall in House Prices < 10% 2.04% 22.02% 11.37%Initial LTV Ratio < 90% × 10% < Fall in House Prices 0.32% 17.42% 1.23%Initial LTV Ratio > 90% × Fall in House Prices < 5% 15.90% 11.08% 23.48%Initial LTV Ratio > 90% × 5% < Fall in House Prices < 10% 0.22% 6.39% 5.31%Initial LTV Ratio > 90% × 10% < Fall in House Prices 0.04% 5.74% 0.67%
Number Registered 3,877,929 3,856,036 4,048,668
36
Table 5: Participation by Financial Distress
For each group I determine the share of registered voters in that group that participated in each of the national elections between 2008 and 2012. Allvariables are defined as in table 4.
Election
2008 Primary 2008 General 2010 Primary 2010 General 2012 Primary 2012 General
Homeowner’s LTV RatioLTV less than 90% 40.43% 79.36% 12.41% 46.70% 46.50% 84.10%LTV greater than 90% 35.88% 74.71% 7.42% 34.09% 40.76% 81.60%
Zip Code Home Price GrowthFall in House Prices < 5% 38.52% 71.29% 12.02% 39.70% 41.68% 75.03%5% < Fall in House Prices < 10% 34.50% 71.27% 11.03% 37.54% 38.29% 72.13%10% < Fall in House Prices 35.12% 69.97% 12.24% 37.25% 33.25% 71.16%
Homeowner’s Treatment GroupInitial LTV Ratio < 90% × HP Fall < 5% 40.18% 79.21% 11.71% 45.52% 47.50% 84.94%Initial LTV Ratio < 90% × 5% < HP Fall < 10% 28.26% 75.64% 10.61% 44.13% 45.73% 82.02%Initial LTV Ratio < 90% × 10% < HP Fall 34.31% 73.36% 11.15% 43.44% 42.07% 79.07%Initial LTV Ratio > 90% × HP Fall < 5% 37.70% 75.37% 8.56% 38.79% 42.53% 81.67%Initial LTV Ratio > 90% × 5% < HP Fall < 10% 28.41% 72.25% 7.07% 37.89% 38.98% 79.91%Initial LTV Ratio > 90% × 10% < HP Fall 31.00% 67.05% 7.86% 37.01% 38.18% 80.68%
37
Table 6: The Effects of House Price Drops on High and Low Leveraged Homeowners
The dependent variable is a dummy variable equal to 100 if the voter participated in the election. The sample consists of all homeowners registered tovote in the 2008 through 2012 general elections and their primaries who moved in or last refinanced after 2003, who were eligible to vote in the 2008general election, whose predicted 2011 home value is not more than 50% different than its assessed value, who do not have an outstanding HELOC, andwho have not refinanced in the previous 2 years. Control variables are dummies equal to 1 if they participated in the 2008 midterm election, participatedin the 2008 general election, are white, are Hispanic, are male or not, and were born in North Carolina. Also in the controls are their birth year cohort,move-in year, house price value two years previously, and the percent change of the median home value between the start of 2003 and the end of 2007.All variables are defined as in tables 2 and 4. Standard errors, adjusted for clustering at the ZIP code level, are reported in parentheses. *, **, and ***denote statistical significance at the 10%, 5%, and 1% level, respectively.
Dependent Variable Voted in the Election (=100)
(1) (2) (3) (4) (5) (6)
Initial LTV Ratio > 90% × % Drop in House Prices -0.151*** -0.134*** -0.0754** -0.0725** -0.0760** -0.0722**(0.039) (0.037) (0.031) (0.031) (0.031) (0.031)
Initial LTV Ratio > 90% 0.144 0.0322 0.142 0.145 0.213 0.198(0.249) (0.250) (0.234) (0.234) (0.228) (0.226)
% Drop in House Prices -0.00399 -0.00972 -0.0195 -0.00959(0.042) (0.041) (0.045) (0.046)
Control Variables YES YES YES YES YES YES
Fixed EffectsElection YESParty-by-Election YES YES YES YES YESCounty-by-Election YES YESInitial House Price Level-by-Election YES YESZip Code-by-Election YES YES
N 740,060 740,060 740,048 740,048 740,054 740,054Adjusted R-Squared 0.379 0.387 0.393 0.393 0.396 0.396
38
Table 7: Comparing Renters, Low-LTV Homeowners, and High-LTV Homeowners
The dependent variable is a dummy variable equal to 100 if the voter participated in the election. The sample is the same as that described in table6 except I now include renters. I consider the effects of house price declines on three groups: renters (the ommitted group), homeowners with initialLTV ratios below 90%, and homeowners with LTV ratios above 90%. Control variables are dummies equal to 1 if they participated in the 2008 midtermelection, participated in the 2008 general election, are white, are Hispanic, are male or not, and were born in North Carolina. Also in the controls aretheir birth year cohort and the percent change of the median home value between the start of 2003 and the end of 2007. All variables are defined as intables 2 and 4. Standard errors, adjusted for clustering at the ZIP code level, are reported in parentheses. *, **, and *** denote statistical significanceat the 10%, 5%, and 1% level, respectively.
Dependent Variable Voted in the Election (=100)
(1) (2) (3) (4)
LTV > 90% Homeowner × % Drop in House Prices 0.0086 -0.0955*** -0.0523** -0.0522**(0.0318) (0.0270) (0.0229) (0.0221)
LTV < 90% Homeowner × % Drop in House Prices 0.0952*** -0.0491*** -0.0341** -0.0367**(0.0193) (0.0178) (0.0153) (0.0151)
LTV > 90% Homeowner 5.147*** 4.903*** 4.828*** 4.906***(0.1560) (0.1520) (0.1340) (0.1270)
LTV < 90% Homeowner 5.415*** 5.371*** 5.124*** 5.196***(0.1360) (0.1330) (0.1120) (0.1010)
% Drop in House Prices -0.0589** -0.0378* -0.0355(0.0260) (0.0210) (0.0244)
Control Variables YES YES YES YES
Fixed EffectsElection YESParty-by-Election YES YES YESCounty-by-Election YESZip Code-by-Election YES
N 10,762,387 10,762,387 10,762,385 10,762,387Adjusted R-Squared 0.445 0.459 0.463 0.465
39
Table 8: Individual Voter Fixed Effects
The dependent variable is a dummy variable equal to 100 if the voter participated in the election. The sample consists of all homeowners registered tovote in the 2008 through 2012 general elections and their primaries who moved in or last refinanced after 2003, who were eligible to vote in the 2008general election, whose predicted 2011 home value is not more than 50% different than its assessed value, who do not have an outstanding HELOC, andwho have not refinanced in the previous 2 years. All variables are defined as in tables 2 and 4. Standard errors, adjusted for clustering at the ZIP codelevel, are reported in parentheses. *, **, and *** denote statistical significance at the 10%, 5%, and 1% level, respectively.
Dependent Variable Voted in the Election (=100)
(1) (2) (3) (4) (5) (6)
Initial LTV Ratio > 90% × % Drop in House Prices -0.323*** -0.309*** -0.203*** -0.199*** -0.189*** -0.186***(0.050) (0.047) (0.038) (0.037) (0.037) (0.037)
Initial LTV Ratio > 90% -2.520*** -2.516*** -2.517*** -2.506*** -2.557*** -2.550***(0.474) (0.447) (0.417) (0.412) (0.411) (0.409)
% Drop in House Prices 0.1000* 0.0774 -0.0357 -0.019(0.053) (0.053) (0.051) (0.054)
Fixed EffectsIndividual Voter YES YES YES YES YES YESElection YESParty-by-Election YES YES YES YES YESCounty-by-Election YES YESInitial House Price Level-by-Election YES YESZip Code-by-Election YES YES
N 779,752 779,752 779,740 779,740 779,745 779,745Adjusted R-Squared 0.508 0.510 0.514 0.515 0.516 0.517
40
Table 9: Homogeneity Across Political Party Affiliations
The dependent variable is a dummy variable equal to 100 if the voter participated in the election. The sampleconsists of all homeowners registered to vote in the 2008 through 2012 general elections and their primarieswho moved in or last refinanced after 2003, who were eligible to vote in the 2008 general election, whose pre-dicted 2011 home value is not more than 50% different than its assessed value, who do not have an outstandingHELOC, and who have not refinanced in the previous 2 years. Control variables are dummies equal to 1 if theyparticipated in the 2008 midterm election, participated in the 2008 general election, are white, are Hispanic,are male or not, and were born in North Carolina. Also in the controls are their birth year cohort, move-inyear, house price value two years previously, and the percent change of the median home value between thestart of 2003 and the end of 2007. All variables are defined as in tables 2 and 4. Standard errors, adjusted forclustering at the ZIP code level, are reported in parentheses. *, **, and *** denote statistical significance atthe 10%, 5%, and 1% level, respectively.
Dependent Variable Voted in the Election (=100)
(1) (2) (3) (4)
LTV > 90% × % Drop HP × Democrat 0.00109 0.0345 0.0348 0.0443(0.0725) (0.0721) (0.0724) (0.0716)
LTV > 90% × % Drop HP × Republican 0.0652 0.0590 0.0578 0.0687(0.0632) (0.0645) (0.0646) (0.0635)
Main Effects and Interaction Effects YES YES YES YESControl Variables YES YES YES YES
Fixed EffectsElection YESCounty-by-Election YES YESInitial House Price Level-by-Election YESZip Code-by-Election YES
N 740,060 740,048 740,048 740,054Adjusted R-Squared 0.385 0.391 0.391 0.394
41
Table 10: Participation Likelihood as a Function of Current LTV Ratio
The dependent variable is a dummy variable equal to 100 if the voter participated in the election. I use the homeowner’s current LTV to determine ifthey are equity rich (LTV < 80%), have low equity (80% < LTV < 100%), or are underwater (current LTV > 100%). The sample consists of all homeownersregistered to vote in the 2008 through 2012 general elections and their primaries who moved in or last refinanced after 2003, who were eligible to votein the 2008 general election, whose predicted 2011 home value is not more than 50% different than its assessed value, who do not have an outstandingHELOC, and who have not refinanced in the previous 2 years. All variables are defined as in tables 2 and 4. Standard errors, adjusted for clustering atthe ZIP code level, are reported in parentheses. *, **, and *** denote statistical significance at the 10%, 5%, and 1% level, respectively.
Dependent Variable Voted in the Election (=100)
(1) (2) (3) (4) (5) (6)
Low Equity -2.292*** -2.233*** -1.733*** -1.577*** -2.023*** -1.709***(0.375) (0.371) (0.337) (0.330) (0.419) (0.346)
Underwater -4.831*** -4.734*** -3.395*** -2.817*** -3.283*** -3.166***(0.593) (0.575) (0.500) (0.491) (0.636) (0.524)
Fixed EffectsIndividual Voter YES YES YES YES YES YESElection YESParty-by-Election YES YES YES YES YESCounty-by-Election YES YESZip Code-by-Election YESCensus Block-by-Election YESInitial House Price Level-by-Election YES
N 852,658 852,658 852,642 852,600 789,382 852,642Adjusted R-Squared 0.507 0.510 0.514 0.516 0.502 0.514
42
Table 11: Financial Distress Groups by Party Affiliation
For each homeowner in the sample, I determine if they have a low initial LTV ratio (LTV < 90%) or high LTVratio (LTV > 90%). I then break them out by party affiliation and their change in house price over the previoustwo years. Below is the share in each of the 18 categories of party by initial leverage by change in house pricesover the panel.
Initial LTV Ratio < 90% Initial LTV Ratio > 90%
< 5% 5% - 10% > 10% < 5% 5% - 10% > 10%
Independent 56.16% 13.62% 6.95% 15.83% 4.92% 2.52%Democrat 55.18% 14.30% 6.93% 15.49% 5.36% 2.73%Republican 58.29% 13.34% 6.80% 15.08% 4.34% 2.14%
43
Table 12: North Carolina County Level Results
The first three models predict the participation rate of households affiliated with no party, with the Democratic party, and with the Republican party,respectively. Models 4 and 5 predict the share of the county vote that went to the the Democratic candidate for president and the share that went to theRepublican candidate for president. Included in the sample are 46 counties in North Carolina at each of the 2008, 2012, and 2016 general elections. Thepercent drop in house prices measures the change in house price of the county’s median home in the two years leading up to the election. % of CountyDem and % of County Rep measure the percent of the counties registered voters that are affiliated with the Democratic party and Republican party,respectively.
Dependent VariableParticipation Rates of County Vote Share for
Independents Democrats Republicans Dem Candidate Rep Candidate
% Drop in County House Prices -0.254*** -0.130*** -0.0755* 0.0243 0.0344(0.058) (0.038) (0.040) (0.019) (0.021)
% of County Dem -2.481*** -1.658*** -3.187*** 0.900*** -0.0723(0.377) (0.246) (0.260) (0.124) (0.135)
% of County Rep -1.025* -0.358 -2.199*** -0.748*** 1.541***(0.607) (0.397) (0.419) (0.200) (0.218)
County FEs YES YES YES YES YES
N 124 124 124 124 124Adjusted R-Squared 0.69 0.738 0.79 0.99 0.988
44
Table 13: Effects of Financial Distress on Homeowners Faraway from their Nearest Polling Place
The dependent variable is a dummy variable equal to 100 if the voter participated in the election. Faraway polling place is a dummy equal to 1 if thevoter’s closest polling place is more than one mile away from their home and zero otherwise. The sample consists of all homeowners registered to vote inthe 2008 through 2012 general elections and their primaries who moved in or last refinanced after 2003, who were eligible to vote in the 2008 generalelection, whose predicted 2011 home value is not more than 50% different than its assessed value, who do not have an outstanding HELOC, and whohave not refinanced in the previous 2 years. Control variables are dummies equal to 1 if they participated in the 2008 midterm election, participated inthe 2008 general election, are white, are Hispanic, are male or not, and were born in North Carolina. Also in the controls are their birth year cohort,move-in year, house price value two years previously, and the percent change of the median home value between the start of 2003 and the end of 2007.All variables are defined as in tables 2 and 4. Standard errors, adjusted for clustering at the ZIP code level, are reported in parentheses. *, **, and ***denote statistical significance at the 10%, 5%, and 1% level, respectively.
Dependent Variable Voted in the Election (=100)
(1) (2) (3) (4) (5) (6) (7) (8)
LTV > 90% × % Drop HP × Faraway Polling Place -0.203*** -0.211*** -0.188*** -0.195*** -0.163** -0.170** -0.160** -0.165**(0.073) (0.071) (0.068) (0.068) (0.068) (0.068) (0.069) (0.069)
Main Effects and Interaction Effects YES YES YES YES YES YES YES YESControl Variables YES YES YES YES YES YES YES YES
Fixed EffectsElection YESParty-by-Election YES YES YES YES YES YES YESCounty-by-Election YES YESInitial House Price Level-by-Election YES YES YESZip Code-by-Election YES YESPolling Place-by-Election YES YES
N 466,539 466,539 466,539 466,539 466,530 466,530 466,442 466,442Adjusted R-Squared 0.384 0.392 0.397 0.397 0.400 0.401 0.403 0.404
45
A Voting
A.1 Theories of Voting
Trying to understand why people vote has been a puzzle for a long time (Aldrich, 1993; Downs, 1957). Po-
tential voters weigh the cost of voting, which they must pay to participate, against the benefit, which is some
combination of the utility they receive if their candidate wins and the chance their vote affects the outcome.
If their preferred candidate was going to win anyway, they would have gotten the benefit without paying the
cost. And if their preferred candidate was going to lose anyway, then they were not going to get the benefit
regardless and paying the cost was a waste. In any election with lots of eligible voters, it would seem that the
chance a given vote might change the outcome is vanishingly small. Therefore, we would expect to see only a
small number of participants. But this flies in the face of the fact that elections are very highly subscribed. A
paradox.
The model just described is called the rational choice model and takes three inputs. The first, B, is the
utility the voter derives if her preferred candidate is elected. The second is the fixed cost, C, she must pay to
participate. The third is the probability that her vote is pivotal, colloquially the probability that her vote is not
wasted, which I call p. The voter then participates if and only if her payoff, pB−C, is positive, and abstains
otherwise. Trying to reconcile the high participation rates observed in elections with the theory, which requires
either p or B to be much larger than most researchers believe them to be, has proved challenging. Many
attempts have been made to fix the theory (Blais, 2000).
The first strategy has been to adjust the rational choice model. Riker and Ordeshook (1968) introduce
a parameter D, duty, that represents the positive utility people derive just from participating, regardless of
the outcome. According to this model, voters participate when pB−C+D is positive. The concern here is a
theoretical one. While it fits nicely with survey results which consistently show that people say they vote be-
cause they feel socially and morally obligated to, the model now begs the question. That is, if it concludes that
people vote because they have a preference for voting, the model has proved itself tautological and ultimately
completely unhelpful. That said, adding D and even I, which represents a voter’s inherent proclivity to be
politically engaged, to the rational choice model is a popular tactic.
Many other adjustments to the model have also been considered. Downs (1957) presented the first solution
to the paradox and suggested that people vote, not just because they prefer some outcome, but because they
perceive there to be a chance that not voting might lead to a utility decreasing collapse of Democracy. Other
papers include regret avoidance, which drops p (Ferejohn and Fiorina, 1974), incorporate game theoretical
elements such that people assume everybody else will also abstain leading them to ex ante believe p is quite
high (Mueller, 1989), and suggest voters consider the disutility of telling people they did not vote (Niemi,
46
1976). Verba and Nie (1972) argue that the decision to vote is driven entirely by the resources, especially time,
that the voter has available to her. Each new model has its limitations. Some researchers wonder if perhaps
the rational voter choice model should be scrapped entirely. For example, Aldrich (1993) suggests that the
costs and benefits to voting are both so low and p so hard to determine that for no voter is it profitable to try
solve his participation problem.
A variety of work has agreed with this assessment. Rosenstone et al. (1993) writes a model in which social
networks, not individuals, make decisions and mobilize all members of the network. By making decisions in
groups, the p and B parameters both carry more weight. Fowler (2006) takes a sociological view and says that
voters think not just about themselves but rather consider the benefits to everybody should their preferred
candidate win. Plutzer (2002) and Gerber et al. (2003) suggest that people’s choices are driven not by some
rational calculus but rather by forces of habit and socialization. Other models consider the role of cognitive
characteristics and psychologies (Smets and Van Ham, 2013). Still other theoretical work posits that thinking
about individual voters is the wrong way to try to understand voting and we should instead focus on contextual
factors (Blais, 2006; Cancela and Geys, 2016). Political institutions and settings, like compulsory voting, close
elections, proportional representation, and the voting age surely all matter (Jackman, 1987). In short, the
research community has yet to agree on nearly anything about what the underlying model of voting might
look like (Geys, 2006).
A.2 Financial Distress Predictions
Following Smets and Van Ham (2013), I consider six categories of theoretical models: rational choice, resource,
mobilization, socialization, psychological, and political institutions. What do these models say about the role
of household financial distress on voter turnout?
Rational choice models can be used to show that financial distress affects voter turnout in multiple ways.
It might be that, as households go from financially secure to financially distressed, they view the role of
government differently and consequently become less indifferent between the candidates. In other words,
financial distress increases B and therefore turnout. Alternatively, B might decrease if financially distressed
voters are more inclined to believe the government is ineffective and unable to affect them. The predicted
effects of how financial distress operates through C are unambiguous. As households become more financially
distressed, the opportunity costs of their time increase, and they are consequently less likely to participate.
Financial distress might also affect D if it causes households to believe the system is broken and not worth
their time. Alternatively, D might increase if voters view their vote as more important. In all, the rational
choice model predicts that financial distress might either increase or decrease participation.
47
Resource models speak most directly to how household financial distress affects turnout. These models
predict that resource constrained households are less likely to participate. Therefore, households experienc-
ing financial distress are unambiguously predicted to be less likely to participate in elections. When viewed
through the lens of the rational choice model, though, the predictions of the resource model become unhelpful.
The resource model is simply the rational choice model without a benefit component. Therefore, using the
resource model instead of the rational choice model to make predictions about the effects of financial distress
is ultimately unhelpful since both models make identical predictions.
Mobilization models argue that people are driven to the polls by social networks that, especially, reduce
the costs of participation. My reading of the literature, along with my own work on the importance of social
networks (McCartney and Shah, 2016), suggests to me that social networks, by better coordinating the actions
of those in the network, can consider their aggregate p and B. As an example, look to the teacher’s union.
By coordinating the actions of teachers across the country, the union’s ability to be pivotal is large. How
mobilization might be affected by financial distress, though, is ambiguous. It might be that as people become
financially distressed, they are less likely to be swayed by their networks, who they feel have failed them.
Or that, as financial distress increases, the mechanisms behind mobilization break down as the organizers
themselves become resource constrained. On the other hand, it might be that some households are happy to
behave independently, but when they become financially distressed feel more strongly that coordination can
be beneficial. Mobilization models, in short, do not make clear predictions about the relationship between
financial distress and participation.
Socialization models suggest that participation decisions are the result of a lifetime of learned behaviors.
A world in which participation decisions are driven just by socialization would see participation unaffected in
the short term by financial distress.
Psychological models might operate, in some sense, through any of the parameters of the rational choice
model. It might be that psychological distress makes it harder for people to solve their voting problem because
they struggle to calculate B, D, or p (Mani et al., 2013). It might be that psychological distress causes depres-
sion which increases C (Lipsman and Lozano, 2011). In both cases, I assume a link between financial distress
and psychological distress. If there is no link, the psychological models predict no effect of financial distress.
Political institutions models suggest that individual determinants are largely irrelevant. Therefore, these
models imply no role for financial distress.
The hope was that this discussion might have shed some theoretical light on how financial distress might
affect turnout. But the waters of prediction remain incredibly murky. The six models make different predic-
tions, and some even make multiple, conflicting predictions, about the effects of financial distress on voter
participation. While the literature struggles with models of voter choice, good empirical work can play an
48
important role by helping guide theory with cleanly identified, causal relationships. Each piece of empirical
evidence helps complete the puzzle as various models better fit with the new finding and the role of various
channels become better understood. It is with this goal in mind that I hope to solve the identification challenge
clouding the relationship between financial distress and participation.
49